Newsletter · · Ashutosh Agarwal

The Moat Question Into Nvidia's Print - AI Accelerators - Week of August 20, 2026

AI accelerators newsletter for the week of August 13-20, 2026. Ben Thompson argued Nvidia's CUDA moat is dramatically diminished as Google pulled Marvell into its custom-silicon camp with a 12.2 billion dollar stock warrant, Michael Burry called Nvidia's 500 billion dollar financing structure shades of Enron, and the AI trade sold off for four straight sessions six days before Nvidia's August 26 print.

AI Accelerators

Week of August 20, 2026: The Moat Question Into Nvidia's Print

GPUs, Custom Silicon and Optics, a twice-weekly read of what the podcasts are actually saying. Issue 017. Window: August 13-20, 2026. Nvidia reports Wednesday, August 26.

On Monday I wrote that the credit market had blinked, that after the roar of the earnings prints, the first serious pushback on the AI buildout had arrived through the debt side of the ledger. Three days later, the doubt has jumped tracks. It is no longer just about whether the money is safe. This week the podcasts started asking a harder question: whether Nvidia's moat is as deep as everyone has priced it to be.

Two things happened at once. Google widened its custom-chip push, this time roping in Marvell alongside its longtime partner Broadcom, and one of the most careful analysts in tech, Ben Thompson, went on a widely followed investing podcast and said out loud that Nvidia's famous software lock-in "is dramatically diminished." Meanwhile the actual stock tape did something it hadn't done in a while: chip and data-center names sold off hard for four straight sessions, one AI networking stock fell 13% in a day, and South Korea's big memory maker announced a $29 billion buyback to stop its own shares from bleeding.

None of this means the buildout is over. Demand still outruns supply, the prints were still enormous, and every skeptic this week said some version of "I'm not calling the top." But the character of the conversation has changed. For two years the debate was "how big." Now, six days before Nvidia's earnings, a growing group of smart people is debating how durable it all is: durable margins, durable moat, durable financing. That is a different and more dangerous question, and it is worth walking through carefully.

TL;DR

  • The moat question is the new story. On Invest Like the Best (Aug 18), Stratechery's Ben Thompson argued the hyperscalers are Nvidia's real threat because they sell their own chips "as commodities," have a lower cost of capital, and because "CUDA's moat is dramatically diminished, the models don't care what they run on." He reads Nvidia's giant financing deal as a defensive response to that.
  • Google pulled Marvell into its chip camp. Per Bloomberg Tech (Aug 19), Google is giving Marvell the right to buy up to $12.2 billion of Google stock as part of a broad chip-design partnership, including an AI accelerator, "the type of chip Nvidia makes." On Squawk on the Street (Aug 19), the warrant detail: 58.97 million shares at a $206.58 strike, already well in the money with Marvell near $241.
  • The financing deal picked up an ugly nickname. On The Accounting Podcast (Aug 19), the hosts relayed that Michael Burry, the investor who called the 2008 crash, says Nvidia's ~$500B funding structure has "shades of Enron" and that he has increased his short positions against AI names.
  • The market actually blinked. Stock Market Today With IBD (Aug 18) called it an "ugly AI sell-off": Nasdaq -1.3%, a "four-day losing streak" in data-center names, Credo -13% in a session, and long-bond yields at a 19-year high because "AI-related borrowing has just exploded."
  • We finally got a real optics deep-dive. The Circuit (Aug 17) walked through Lumentum's strong quarter vs. Coherent's margin stumble, and why the "holy grail" of co-packaged optics is further off than the Street thinks.
  • A fresh operator on inference chips. d-Matrix CEO Sid Sheth (Eye On A.I., Aug 17) described a "premium token economy" where customers pay $20 per million tokens for fast, interactive AI vs. $2 for standard, a segment GPUs are ill-suited for.
  • Negative space: still no Broadcom or Marvell executive on a podcast, no fresh Nvidia roadmap voice, and AMD went essentially silent this week after last issue's flood of operator color.

The developments that matter

1) The moat, not just the money: Ben Thompson makes the bear case that actually bites

The single most important 40 minutes of podcast this week was Ben Thompson, the analyst behind Stratechery, on Invest Like the Best (Aug 18). What makes it matter is that Thompson is not a permabear; he spends much of the episode explaining why he thinks AI's economic impact "is going to be astronomical." His worry is narrower and sharper, and it lands right on Nvidia.

Start with the competitive point. The hyperscalers, Google, Amazon, Microsoft, Meta, are building their own chips, and Thompson thinks that is Nvidia's deepest problem, not a sideshow:

"Google already made a deal to sell like 20% of their TPUs to Anthropic. On the last earnings call, Andy Jassy practically confirmed that they'll be selling Trainium chips externally... And by the way, they're not selling their chips on differentiation. They're selling their chips as commodities. Nvidia is the one selling on differentiation."

(Quick definitions, because the jargon hides the point. A TPU is Google's in-house AI chip; Trainium is Amazon's. CUDA is Nvidia's software layer, the thing that historically made it painful to switch away from Nvidia hardware. A hyperscaler is a giant cloud owner like Google or Amazon. A NeoCloud is a newer company like CoreWeave or Nebius that does nothing but rent out AI compute.)

Here is the line that will get quoted for weeks:

"CUDA's moat is dramatically diminished because the models don't care what they run on."

His logic: the value is increasingly in the AI models and what's built on top of them, not in the specific chip underneath. And the hyperscalers have the one thing Nvidia's other customers don't, "scale, lower cost of capital. It's a capital fight." Even Elon Musk's loudly pro-Nvidia stance gets reframed: SpaceX buys Nvidia, Thompson says, not "because they're the best" but "because they're the most fungible", the easiest hardware to rent back out when you're not using it.

Then he connects it to the deal everyone's talking about. Thompson reads Nvidia's ~$500 billion financing platform as a reaction to the Google threat:

"That's how you get this deal this week. I see this deal as a response... it goes with the Google deal. Google can just issue equity. The shareholders don't love it, but their monetization capacity is much higher than Nvidia's customers are."

Why it matters for the thesis: For two years the bull case rested partly on CUDA as a durable moat and on Nvidia's ~80% gross margins as a sign of pricing power that would last. If the moat is eroding and the biggest customers are also the most credible competitors, then the terminal-margin assumption, the number that justifies a $4-5 trillion market cap, is the thing at risk, not the next quarter's revenue. This is the same conclusion Vincent Daniel reached last issue from the margin side (80% margins drifting "to 60 or 40"); Thompson gets there from the moat side. When two independent thinkers arrive at the same place from different doors, pay attention.

Thompson also left two gifts for anyone building a long-term model. First, on memory makers, he warned they may have overplayed their hand, comparing them to Iran closing the Strait of Hormuz: devastating once, but "no one's going to let themselves get in this situation again," with Apple lobbying for Chinese memory and every AI lab now optimizing to use less memory. Second, on what survives if this is a bubble: "The GPUs don't last that long... What's going to last from AI? It has to be power." That is a tell for where the durable value may end up, not in the silicon, but in the electricity.

2) Google + Marvell: the custom-silicon camp gets a second general

The concrete event behind Thompson's thesis landed midweek. On Bloomberg Tech (Aug 19), semiconductor reporter Ian King explained that Google is giving Marvell the right to buy as much as $12.2 billion of Google stock, wrapped inside a broad chip-design partnership. Most of the products, King said, are "the backup singers in the band", but one is not:

"There is one important thing in there, which is the AI accelerator. This is the type of chip that Nvidia makes. It's the cornerstone of Nvidia's product... this shows that Google is becoming more ambitious about its own chip efforts and is looking to use Marvell to help it."

King's framing of the industry structure is the useful part: Marvell and Broadcom "are the absolute leaders in this market" where a company brings its own bright idea and IP, and these two "package it all together and take it off to TSMC and get it manufactured for you." Google already does this with Broadcom for the TPU. Adding Marvell is a deliberate move to second-source its custom silicon, exactly the multi-vendor discipline you'd expect from a buyer trying to reduce dependence on any one supplier (including Nvidia).

On Squawk on the Street (Aug 19), the desk added the warrant mechanics: Google can buy up to 58.97 million Marvell shares at a $206.58 strike; with Marvell trading around $241, that's already deep in the money. The hosts stressed the TPU "was made by Broadcom, so Broadcom did not lose the contract; it augments the contract," and that the new Marvell piece adds a memory-augmentation capability ("and we know how desperately everyone wants memory"). One host got visibly excited, calling it "a revolutionary deal" and pointing at a "$120 billion" opportunity against Marvell's ~$11 billion of revenue, but to be clear, that $120 billion is his own enthusiastic characterization of the potential, not a disclosed contract value. What is disclosed is the $12.2 billion stock warrant. Marvell's market cap has cleared $200 billion, and the name is up ~180% year-to-date.

Why it matters: This is the custom-silicon theme graduating from "narrative" to "signed paper." And notice the same circular-financing question that dogs Nvidia now attaches to the chip designers too, as King put it, "if this is such a good deal, why are we giving something back to our customers?" The AI complex keeps inventing new ways to hand equity or guarantees to its own buyers.

3) "Shades of Enron": the credit doubt hardens

Last issue the credit doubt arrived as a downgrade. This week it arrived as an accusation. On The Accounting Podcast (Aug 19), hosts Blake Oliver and David Leary, accountants by trade, which is the point, relayed that Michael Burry (yes, the Big Short investor) has publicly called Nvidia's ~$500 billion funding structure a case of "shades of Enron," and has increased his short positions against major AI companies.

The mechanism they walked through is worth understanding in plain terms. A group of financial giants (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR) sets up special purpose vehicles (separate legal entities, "SPVs") that borrow money to build data centers and buy Nvidia chips. Those SPVs then lease the computing power to the likes of OpenAI. Nvidia guarantees roughly 25% of the debt. The hosts' key insight is why Nvidia does it this way rather than building the data centers itself:

"They could build their own data centers if they wanted to, but they're not doing it. Why? Because they want to actually book the revenue by selling the chips to these other entities. If they built the data centers themselves... they'd have all these expenses and all this capital."

That's the heart of Burry's Enron comparison, he called the structure an attempt to use "unnatural credits to prolong momentum late in the bull phase," a "sign of desperation." (For the too-young-to-remember: Enron used off-balance-sheet entities to hide debt and pull revenue forward; its ~$60 billion collapse in 2001 was then the largest US bankruptcy and is why Sarbanes-Oxley and the accounting-oversight board exist.) The accountants were careful and fair: "There's no fraud happening here. This is all happening in plain sight... maybe the problem is that GAAP allows this." They also flagged the depreciation game underneath it all, AI chips have a real useful life of "two to three years," but companies have been stretching that to "five or six years" on paper to report lower depreciation and higher profit.

Two more human sources fill in the numbers. On Facts vs Feelings (Aug 19), Carson's Grant Barry gave the cleanest timeline anyone has laid out. On July 22, Nvidia said it would backstop ~$250 billion for OpenAI to lease compute plus ~$350 billion of chip-purchase financing, and its five-year credit default swaps (the market's price for insuring against a Nvidia default) "nearly doubled," hitting 82 basis points in late July, up from about half that earlier in the year. Then on August 11, Jensen announced the $500 billion memorandum of understanding ("nothing's been committed," Barry stressed) with BlackRock, Blackstone, Brookfield, Goldman and KKR, of which, he noted drily, "technically, Goldman's the only bank; the rest are private capital and asset managers." Nvidia's role is a residual-value backstop of up to 25%, a ceiling around $125 billion, that "kicks in only after the NeoCloud goes under." His verdict is the sentence to keep:

"The tail risk here is being taken off Nvidia's balance sheet... but their tail risk hasn't gone away. It's transferring to the rest of the market, to all of us."

After the August 11 news, Barry noted, Nvidia's CDS eased back to about 70 basis points and the stock, which had fallen ~10% in the last week of July, gained ~18%. So the credit market's immediate reaction was relief that Nvidia offloaded risk, even as that risk simply moved onto the institutions and pension savers buying the SPV debt. He added that Meta's and Alphabet's CDS have also risen, and that Oracle's is "the worrisome one."

The details of the OpenAI piece came from a working journalist. On The Information's TITV (Aug 17), Nvidia reporter Phoebe Liu reported that Nvidia's guarantee on the OpenAI/SB Energy 10-gigawatt Ohio project totals about $105 billion of maximum exposure (per a regulatory filing), "not the amount Nvidia is paying," and a number that declines over time. Crucially, this is softer than Nvidia's earlier CoreWeave guarantee (a promise to buy back 100% of unsold GPU capacity), which Liu called "more airtight." She reported the guarantee may have been phased rather than cut, the first tranche covers 4.5 gigawatts, with a second 3.75-gigawatt phase Nvidia is "technically not on the hook for yet", which she read as Nvidia deliberately "signaling financial discipline" and capping its own exposure. Nvidia is also putting $1.5 billion (possibly $3 billion) into SB Energy itself. Liu's closing line is the honest one: Jensen thinks the odds Nvidia ever pays are "close to zero" because demand outpaces supply, "but bond traders' assessment of Nvidia's credit risk has gone up over the last few months."

Why it matters: This is no longer one bank's downgrade. It's a famous short-seller, a set of accountants, a credit strategist, and an investigative reporter all circling the same structure in the same week, and a measurable move in Nvidia's own credit-insurance price. The bull retort ("demand outpaces supply, the backstop never triggers") is real and may well be right. But the burden of proof has shifted, and it shifts further every time a new counterparty gets added to the daisy chain.

4) The market didn't just talk about it, it sold

Underneath the debate, the tape finally cracked. On Stock Market Today With IBD (Aug 18), hosts Alissa Coram and Ed Carey described "a lot of damage in the AI trade": the Nasdaq Composite -1.3%, the Nasdaq 100 -1.7%, data-center names on a "four-day losing streak," and what Carey called "a rug pull", stocks that had just broken out getting hit for 10% in a day. The poster child was AI-networking chipmaker Credo, down 13% in a single session the day after an attempted breakout. Carey's advice was blunt: "keep your AI exposure limited until there's more of a foothold... this is not April, May."

He tied the sell-off to something structural, not a one-off headline: long-term interest rates. The 30-year Treasury yield hit a fresh 19-year high and the 10-year matched a 19-month high, and his explanation was pure AI-buildout read-through:

"AI investment-grade debt, AI-related borrowing has just exploded. So there's a lot of demand for money, and that's pushing up yields, public and private... a headwind for the market."

On the Squawk desk the same day, the traders pointed at "a gigantic hit in Korea that really killed us yesterday", a reference to the memory makers, plus fresh data-center resistance in Pennsylvania, a state "supposed to be very data center friendly." The tonic that steadied things Wednesday was the Google-Marvell deal plus a buyback: on Bloomberg Tech, SK Hynix announced a $29 billion share repurchase "to stabilize its stock after falling more than 50% in two months."

Why it matters: Last issue I argued the credit market had blinked. This week the equity market blinked too, a genuine multi-day drawdown, not a wobble, with the same fingerprints on it: too much borrowing, rising rates, and local pushback on power and land. It lands six days before Nvidia reports, which raises the stakes on that print considerably.

5) The optics deep-dive we've been missing

For weeks the optics story (the transceivers and fiber links that shuttle data between GPUs) showed up only as earnings-recap noise. This week The Circuit (Aug 17), Ben Bajarin and Jay Goldberg, finally did the technical work, and it's the most useful thing said about optics all month.

The setup: Lumentum "beat across the board, guided up," while Coherent's revenue guide was fine but it "struggled with margins," and the two stocks diverged sharply. The reason is a manufacturing detail most investors gloss over, the move to bigger, six-inch indium phosphide wafers (the exotic material optical chips are made from). Bigger wafers should mean more chips per wafer and fatter margins. It isn't playing out, because:

"Six-inch indium phosphate wafers are brand new. It is not a mature ecosystem... there's only one supplier of six-inch indium phosphate wafers, and they're very new to it too."

Lumentum is doing well squeezing four-inch wafers and reusing that know-how; Coherent, "perceiving it was behind," pushed hard into six-inch and hit the growing pains. Semiconductor economics 101, as Goldberg put it.

The bigger, more valuable point was a cold shower on the "optics is about to explode" trade. The hosts laid out the ladder clearly: today's standard is LPO (a pluggable module at the edge of the board); next comes NPO (near-packaged optics, moved onto the board); and the "holy grail" is CPO (co-packaged optics, fiber attached directly to the chip). Their verdict: CPO is "the wild, wild west," years from scale, and "doesn't scale, the manufacturing bottlenecks here are too deep." NPO, they think, "lasts for a long time." The warning to the Street:

"People just keep getting ahead of their skis... it's a very slow roll. These companies are going to be gigantic, but it's a very slow roll."

They also flagged an under-appreciated tell from Applied Materials' quarter (reported Aug 14): the chip-equipment giant now has "eight quarters of visibility" from customers, versus "two quarters" in prior cycles, evidence that the long-term supply agreements rippling through the supply chain ("LTA waterfalling") are structural, not just hype. And the "unspoken guest" on Applied's call, in Bajarin's read, was Intel, Applied wouldn't give a full-year guide because it doesn't yet know how much capacity Intel will order.

Why it matters: For anyone tempted to chase Coherent, Lumentum, Credo or Astera on the optics story, this is the nuance the headline numbers hide. The demand is real and the long-term agreements are real, but the highest-value technology (CPO) is a slow, bottlenecked grind, and margin outcomes hinge on unglamorous wafer-yield details. Buy the theme if you like it, but not on the assumption that co-packaged optics scales next year, because the people who research it for a living say it won't.

6) A fresh operator: d-Matrix and the "premium token economy"

Amid all the finance talk, one actual chip executive gave genuinely new operating color. On Eye On A.I. (Aug 17), d-Matrix CEO Sid Sheth described a market splitting in two. Ordinary AI inference (running a trained model to answer a prompt) is a commodity, but a new premium tier is exploding, fast, highly interactive uses like AI coding tools and agents, where speed keeps users engaged:

"You can always say I don't want that level of interactivity, in which case I'll pay $2 for a million tokens. But if I want that high level of interactivity, I'll pay $20 for a million tokens. And people are willing to pay that $20."

His pitch is that GPUs are poorly suited to this "premium token economy" because they're bottlenecked by HBM memory bandwidth, whereas architectures like d-Matrix's (and rivals Groq and Cerebras) put "compute and memory together... an order of magnitude faster than HBM." He said the segment "has exploded on us literally in the last few months," and to keep up d-Matrix bought Giga.io to speed rack deployment and is partnering with Supermicro to build racks to spec, selling to frontier labs and sovereign customers.

Why it matters: This is the clearest operator-level evidence yet that inference, not training, is the next battleground, and that a slice of it may not belong to Nvidia at all. Sheth pegs inference as a ">$1 trillion" market within five years. Even if you discount a founder talking his own book, the "$20 vs $2 per million tokens" split is a concrete, checkable claim about where pricing power in AI is actually forming.

7) Where the money is quietly moving: the optics-and-power tilt

Filings season gave a read on positioning, via Limitless: An AI Podcast (Aug 18), worth noting this is an analysis/commentary show, so treat the numbers as reported rather than confirmed. The hosts parsed 13F filings (the quarterly disclosures big investors must file): Berkshire took a large Google position; Brad Gerstner's book is led by Nvidia ($1.9B) with Cerebras, Meta, TSMC, CoreWeave and ARM behind it; and Gavin Baker's largest position is SpaceX ($4.7B), followed by Micron, Cerebras, Astera Labs, Ciena and Credo. The common thread among the aggressive AI investors, the hosts said, is "the power and optics trade", Coherent (up ~85% year-to-date) for the light-based interconnects that are replacing copper between GPUs, and Astera Labs for the "plumbing."

There was also a cautionary tale that fits this week's mood: the blow-up of a heavily memory-tilted book ("Leopold's ghost portfolio", reportedly ~28% SanDisk, ~28% Micron) that was "directionally correct" but "got totally liquidated" on leverage. The lesson the hosts drew is the one this whole issue keeps circling: in a market this reflexive, being right on direction isn't enough if the financing underneath you breaks first.

The debate, steel-manned

Bull case (durable moat, durable demand): Demand genuinely outpaces supply, 13-14x by one estimate cited earlier this month, so payback periods of ~2 years are real (CoreWeave and Nebius both said as much). Nvidia's financing platform is smart vendor financing that widens its customer base beyond a handful of hyperscalers, and Jensen's team is deliberately capping its own risk at 25% and phasing the OpenAI guarantee, signs of discipline, not desperation. On the technology, even Ben Thompson concedes the economic opportunity is "astronomical" and that Nvidia stays "the most token-efficient" in a power-constrained world, which is the world we're heading into. Custom chips from Google and Amazon are commodities sold on price; Nvidia still sells on performance and fungibility, and that's worth a premium.

Bear case (moat and margins erode, financing is the fault line): CUDA's lock-in is fading because models are hardware-agnostic; the hyperscalers are both Nvidia's biggest customers and its most credible competitors, and they have a lower cost of capital. That combination pressures the ~80% gross margin that justifies the valuation. The financing structure pulls revenue forward while pushing tail risk onto pension savers and private-credit investors ("shades of Enron," in Burry's words), and it only works while everyone keeps buying next year's chips at higher prices. The tape is already voting: a four-day sell-off, Credo -13%, SK Hynix -50% in two months, 19-year-high long-bond yields driven by AI borrowing, and rising credit-default-swap prices on Nvidia, Meta, Alphabet and (worst) Oracle.

The swing factor everyone now agrees on: credit and cost of capital. Bulls and bears this week converged on the same dashboard, Nvidia's and Oracle's CDS spreads, hyperscaler and SPV bond issuance, long-bond yields, and whether that $500 billion of third-party money actually shows up. Watch those, not the next revenue headline. As Thompson put it, the nearest-term question may not be "will there be enough demand" but "are we going to have enough money."

Names in play

  • Nvidia (NVDA): The whole issue is about it, and it reports Wednesday, Aug 26. The debate has moved from "how big is the beat" to "how durable are the margins and the moat." Watch the gross-margin line and any commentary on custom-silicon competition and the financing backstop.
  • Marvell (MRVL) & Broadcom (AVGO): The custom-silicon "leaders." Marvell just got the Google warrant deal (+~180% YTD, >$200B cap); Broadcom keeps the TPU contract, which the Marvell deal "augments" rather than replaces. Still no executive from either on a podcast, a coverage gap.
  • Google (GOOGL): The aggressor of the week (Marvell + the existing Broadcom TPU relationship + selling ~20% of TPUs to Anthropic), and, per Thompson, the company Nvidia's financing deal is really a response to.
  • Memory, SK Hynix, Micron (MU), SanDisk: SK Hynix's $29B buyback after a 50% two-month drop is the tell that the memory trade got over its skis. Thompson's "Strait of Hormuz" warning is the long-term bear note; SanDisk's high-bandwidth-flash backlog into 2027 is the bull sub-plot.
  • Optics, Lumentum (LITE), Coherent (COHR), Credo (CRDO), Astera Labs (ALAB): Real demand, real long-term agreements, but co-packaged optics is a slow grind and margins hinge on wafer yields. Credo's 13% single-day drop shows how crowded and jumpy this corner is.
  • Inference challengers, d-Matrix, Groq, Cerebras: The "premium token economy" ($20 vs $2 per million tokens) is where non-Nvidia architectures are trying to carve out a beachhead.
  • The financiers, Apollo, BlackRock, Blackstone, Brookfield, KKR, Goldman: Now structurally part of the story. If the AI buildout is becoming a credit cycle, these are the names holding the risk that left Nvidia's balance sheet.

Read-throughs

  • Rates are now an AI story. The clearest new linkage this week: AI-related borrowing is big enough to push long-term Treasury yields to 19-year highs (per IBD). That's a headwind for every rate-sensitive business, housing especially, and it means the AI trade and the bond market are now joined at the hip. A disappointing Nvidia print could ripple into rates, and vice versa.
  • Custom silicon is a Broadcom/Marvell/TSMC read-through, not just a Nvidia negative. Every hyperscaler chip still gets designed with Broadcom or Marvell and built by TSMC. The more Google, Amazon and OpenAI diversify away from Nvidia, the more volume flows to those three, a barbell worth holding if you believe the custom-silicon shift is real.
  • Power is the lasting asset. Thompson's "what survives a bubble is the power" line, plus fresh data-center resistance in Pennsylvania, points the durable value toward electricity and grid infrastructure rather than the chips themselves, chips depreciate in 2-3 years; substations don't.
  • Inference > training for the next leg. Between d-Matrix's premium-token pitch and SanDisk's inference-focused high-bandwidth flash, the center of gravity is shifting toward serving models cheaply and fast, a different competitive map than the training-cluster land grab that defined the last two years.

What changed since Issue 016 (Monday, Aug 17)

  • The doubt moved from money to moat. Issue 016's lead was the credit market's first pushback (the Broadcom XPU downgrade). Issue 017's lead is that the moat is now in question too, Ben Thompson's "CUDA is dramatically diminished," plus Google pulling Marvell into its chip camp. Custom silicon has gone from "a theme" to "signed warrants + a sharp intellectual case."
  • The credit doubt hardened into a name. "The credit blinks" (016) became "shades of Enron" (017), Michael Burry adding shorts, an accounting-fraud framing, and a clean CDS timeline (82bps to 70bps around the Aug 11 announcement).
  • The market actually sold off. Last issue described the first pushback; this issue describes a genuine four-day drawdown, Credo -13%, SK Hynix -50% over two months, and record-long-end yields.
  • Optics flipped from negative space to a real deep-dive. Issue 016 flagged "no dedicated optics technical episode." This week The Circuit delivered exactly that, LPO/NPO/CPO, indium-phosphide yields, and a "slow roll" warning.
  • AMD went quiet. Issue 016 had a flood of fresh AMD operator color (MI350P, MI455X, Helios). This week: essentially nothing, two dedicated searches came back empty. Negative space, and worth noting into Nvidia's print.
  • The next date is fixed: Nvidia reports Wednesday, August 26. This issue is the last one before it.